CODEWITHSUNDEEP
pythonneural-networkcomputer-visiondeep-learningconv-neural-network
subho
subho

Reputation: 591

Why in convolutional neural network 3 channels when con-vol into 3 channels filter produce only 1 channel

enter image description here

I want to understand why a image with 3 channels like 6*6*3 when convol with 3*3*3 filter produce only 4*4 and not 4*4*3

Upvotes: 0

Views: 1717

Answers (2)

Deniz Beker
Deniz Beker

Reputation: 2174

One way is to think how many 3 x 3 x 3 cubes you may cut and grab from 6 x 6 x 3 rectangle prism.

Also, let's make the question simpler.

Let's say you have 2 x 2 input image patch and you want to do 2 x 2 convolution. How many ways you can fit 2 x 2 convolution to input image patch = 1 x 1 (1 in x direction, 1 in y direction)

Let's extend it: 4 x 2 input image with 2 x 2 conv. Unique convolution count = 3 x 1 (3 in x direction, 1 in y direction etc.)

Let's extend it: 4 x 4 input image with 2 x 2 conv. Unique convolution count = 3 x 3

Let's extend it: 4 x 4 x 2 input image with 2 x 2 x 2 conv. Unique convolution count = 3 x 3 x 1

Let's extend it: 6 x 6 x 3 input image with 3 x 3 x 3 conv. Unique convolution count = 4 x 4 x 1

Upvotes: 1

Mozglubov
Mozglubov

Reputation: 433

When you apply a convolution it is summing across the outputs of your kernel weights multiplied by your input values. In this case you are not doing any sort of padding of your input, which means you are only outputting the values which are fully defined. If you take the dimensions of your input as (x,y,z), then you can see that your kernel is smaller in the x and y dimension, but equal in the z. That means that you can slide the kernel in both x and y directions, producing an output for each location, but in the z direction it has nowhere to slide, so it just produces a single output (which is the sum across all channels).

Upvotes: 1

Related Questions